Preview

PDF, English Download (10MB) | Terms of use

Abstract

Analyses of expression profiles between different phenotypes of cancer patient populations or models led to identifying gene modules, or gene expression signatures, that describe specific biological mechanisms. A signature is often used to describe the same mechanism in different contexts, for example, in large-scale cancer gene expression studies. Over the past, gene expression signatures have become readily available and collected in large signature databases like the MsigDB. Although gene expression signatures represent a critical knowledge base for the analysis of cancer populations, they are often carelessly applied or contaminated by confounding processes, such as proliferation. This might be why only a few gene expression signatures have reached clinical relevance. Therefore, in the first chapter of this thesis, I aim to define a workflow that assesses the quality and translatability of gene expression signatures, thereby enabling the interpretation of their associated biological mechanisms in the context of another cancer study. Additionally, I provide an analytical methodology, to analyze a collection of gene expression signatures that describe a broad range of cancer mechanisms. Additionally, I describe a web application named Roset- taSX that uses this methodology and allows users to analyze public molecular data of more than 11,000 cancer patients or cancer models. Finally, I will show the applicability of my approach by recapitulating of the intrinsic breast cancer subtypes and other molecular characteristics. This first project will lay the methodological foundation for two subsequent analyses. In the second project, I use my approach to evaluate a set of gene expression signatures postulated to describe epithelial-to-mesenchymal, mesenchymal, or stemness phenotypes of cancer populations. Previous studies suggested that cells in the tumor microenvironment contribute to individual signatures in specific cancer types. In this study, I applied my methodological approach to analyze multiple levels of data granularity, including cancer cell line and single cell data, in addition to clinical tumor data. The goal was to highlight the impact of contamination on the largest combined set of mesenchymal signatures investigated to date in this context. This project emphasized the significance of thoroughly evaluating cancer cell content when utilizing these signatures. It also demonstrated how incorrect conclusions about cancer characteristics can be drawn when quality control is not rigorously applied in signature analyses. The final chapter, will apply the methodological approach to evaluate the underdiagnosis of large cell neuroendocrine carcinomas (LCNEC) in a real-world data non-small cell carcinoma (NSCLC) cohort. Although LCNEC was not initially classified as a separate subtype of NSCLC, it has been classified as a separate group in the most recent WHO classification recommendation. The increased recognition of LCNEC over the past has shown an increase in the prevalence of this rare disease, accounting for 1%-3% of all lung cancers. However, today, practical limitations, similarities, and overlap with other NSCLCs are still complicating the diagnosis of LCNEC. Based on a RosettaSX analysis that revealed neuroendocrine differentiation in many patients with NSCLC, I will demonstrate how a machine learning model was used to assess the degree of LCNEC underdiagnosis. In summary, this work presents an integrated approach for evaluating gene expression signatures in depth. The signature analysis framework was applied in two ways: for the in-depth assessment of gene expression signatures and the comprehensive characterization of a patient population. The approach described herein can provide robust findings for gene expression signatures that are easily interpretable and can reveal previously unknown associations between biomarkers and expression phenotypes.